Skip to content

~95% on SimpleQA (e.g. Qwen3.6-27B on a 3090). Supports all local and cloud LLMs (llama.cpp, Ollama, Google, ...). 10+ search engines - arXiv, PubMed, your private documents. Everything Local & Encrypted.

14.8 KB
Raw Blame
"""Tests for benchmark-related database models."""

from datetime import datetime, timedelta, timezone

import pytest
from sqlalchemy import create_engine
from sqlalchemy.exc import IntegrityError
from sqlalchemy.orm import sessionmaker

from local_deep_research.database.models import (
    Base,
    BenchmarkConfig,
    BenchmarkProgress,
    BenchmarkResult,
    BenchmarkRun,
    BenchmarkStatus,
    DatasetType,
)


class TestBenchmarkModels:
    """Test suite for benchmark-related models."""

    @pytest.fixture
    def engine(self):
        """Create an in-memory SQLite database for testing."""
        engine = create_engine("sqlite:///:memory:")
        Base.metadata.create_all(engine)
        yield engine
        engine.dispose()

    @pytest.fixture
    def session(self, engine):
        """Create a database session for testing."""
        Session = sessionmaker(bind=engine)
        session = Session()
        yield session
        session.close()

    def test_benchmark_run_creation(self, session):
        """Test creating a BenchmarkRun."""
        run = BenchmarkRun(
            run_name="GPT-4 vs Llama Comparison",
            config_hash="abc123def456",
            query_hash_list=["hash1", "hash2", "hash3"],
            search_config={"engine": "google", "max_results": 10},
            evaluation_config={"model": "gpt-4", "temperature": 0.0},
            datasets_config={"simpleqa": 100, "browsecomp": 50},
            status=BenchmarkStatus.PENDING,
            total_examples=150,
        )

        session.add(run)
        session.commit()

        # Verify the run
        saved = session.query(BenchmarkRun).first()
        assert saved is not None
        assert saved.run_name == "GPT-4 vs Llama Comparison"
        assert saved.config_hash == "abc123def456"
        assert len(saved.query_hash_list) == 3
        assert saved.total_examples == 150
        assert saved.status == BenchmarkStatus.PENDING

    def test_benchmark_status_progression(self, session):
        """Test status progression of a benchmark run."""
        run = BenchmarkRun(
            config_hash="test123",
            query_hash_list=[],
            search_config={},
            evaluation_config={},
            datasets_config={},
            status=BenchmarkStatus.PENDING,
        )

        session.add(run)
        session.commit()

        # Progress through statuses
        run.status = BenchmarkStatus.IN_PROGRESS
        run.start_time = datetime.now(timezone.utc)
        session.commit()

        assert run.status == BenchmarkStatus.IN_PROGRESS
        assert run.start_time is not None

        # Complete the run
        run.status = BenchmarkStatus.COMPLETED
        run.end_time = datetime.now(timezone.utc)
        run.overall_accuracy = 0.85
        run.processing_rate = 5.2  # examples per minute
        session.commit()

        assert run.status == BenchmarkStatus.COMPLETED
        assert run.end_time is not None
        assert run.overall_accuracy == 0.85

    def test_benchmark_result_creation(self, session):
        """Test creating BenchmarkResult records."""
        # Create a benchmark run first
        run = BenchmarkRun(
            config_hash="test123",
            query_hash_list=["q1", "q2"],
            search_config={},
            evaluation_config={},
            datasets_config={},
            status=BenchmarkStatus.IN_PROGRESS,
        )
        session.add(run)
        session.commit()

        # Create results
        result1 = BenchmarkResult(
            benchmark_run_id=run.id,
            example_id="simpleqa_001",
            query_hash="hash_q1",
            dataset_type=DatasetType.SIMPLEQA,
            research_id="research-uuid-123",
            question="What is the capital of France?",
            correct_answer="Paris",
            response="The capital of France is Paris.",
            extracted_answer="Paris",
            confidence="high",
            processing_time=2.5,
            is_correct=True,
            graded_confidence="high",
        )

        result2 = BenchmarkResult(
            benchmark_run_id=run.id,
            example_id="browsecomp_001",
            query_hash="hash_q2",
            dataset_type=DatasetType.BROWSECOMP,
            research_id="research-uuid-124",
            question="Compare Python and JavaScript",
            correct_answer="Python is interpreted, JavaScript runs in browsers",
            response="Python and JavaScript are both popular languages...",
            extracted_answer="Different use cases",
            confidence="medium",
            processing_time=5.2,
            is_correct=False,
            research_error="Timeout during search",
        )

        session.add_all([result1, result2])
        session.commit()

        # Verify results
        results = session.query(BenchmarkResult).all()
        assert len(results) == 2

        correct_result = (
            session.query(BenchmarkResult).filter_by(is_correct=True).first()
        )
        assert correct_result.extracted_answer == "Paris"
        assert correct_result.dataset_type == DatasetType.SIMPLEQA

        failed_result = (
            session.query(BenchmarkResult).filter_by(is_correct=False).first()
        )
        assert failed_result.research_error == "Timeout during search"

    def test_benchmark_config_management(self, session):
        """Test BenchmarkConfig for saving and reusing configurations."""
        config = BenchmarkConfig(
            name="High Accuracy Config",
            description="Configuration optimized for accuracy over speed",
            config_hash="config_abc123",
            search_config={
                "engines": ["google", "bing", "semantic_scholar"],
                "max_results": 20,
                "timeout": 30,
            },
            evaluation_config={
                "model": "gpt-4",
                "temperature": 0.0,
                "max_retries": 3,
            },
            datasets_config={
                "simpleqa": 200,
                "browsecomp": 100,
                "sample_ratio": 0.1,
            },
            is_default=True,
            is_public=True,
        )

        session.add(config)
        session.commit()

        # Test updating usage stats
        config.usage_count += 1
        config.last_used = datetime.now(timezone.utc)
        config.best_accuracy = 0.92
        config.avg_processing_rate = 4.5
        session.commit()

        # Verify config
        saved = (
            session.query(BenchmarkConfig).filter_by(is_default=True).first()
        )
        assert saved is not None
        assert saved.name == "High Accuracy Config"
        assert saved.usage_count == 1
        assert saved.best_accuracy == 0.92
        assert "semantic_scholar" in saved.search_config["engines"]

    def test_benchmark_progress_tracking(self, session):
        """Test BenchmarkProgress for real-time tracking."""
        # Create a benchmark run
        run = BenchmarkRun(
            config_hash="test123",
            query_hash_list=["q1", "q2", "q3"],
            search_config={},
            evaluation_config={},
            datasets_config={},
            status=BenchmarkStatus.IN_PROGRESS,
            total_examples=100,
        )
        session.add(run)
        session.commit()

        # Add progress updates
        progress1 = BenchmarkProgress(
            benchmark_run_id=run.id,
            completed_examples=25,
            total_examples=100,
            overall_accuracy=0.88,
            dataset_accuracies={"simpleqa": 0.90, "browsecomp": 0.85},
            processing_rate=3.2,
            estimated_completion=datetime.now(timezone.utc)
            + timedelta(minutes=20),
            current_dataset=DatasetType.SIMPLEQA,
            current_example_id="simpleqa_025",
            memory_usage=512.5,
            cpu_usage=45.2,
        )

        progress2 = BenchmarkProgress(
            benchmark_run_id=run.id,
            completed_examples=50,
            total_examples=100,
            overall_accuracy=0.86,
            dataset_accuracies={"simpleqa": 0.89, "browsecomp": 0.83},
            processing_rate=3.5,
            estimated_completion=datetime.now(timezone.utc)
            + timedelta(minutes=15),
            current_dataset=DatasetType.BROWSECOMP,
            current_example_id="browsecomp_010",
        )

        session.add_all([progress1, progress2])
        session.commit()

        # Query progress updates
        progress_updates = (
            session.query(BenchmarkProgress)
            .filter_by(benchmark_run_id=run.id)
            .order_by(BenchmarkProgress.timestamp)
            .all()
        )

        assert len(progress_updates) == 2
        assert progress_updates[0].completed_examples == 25
        assert progress_updates[1].completed_examples == 50
        assert (
            progress_updates[1].processing_rate
            > progress_updates[0].processing_rate
        )

    def test_benchmark_relationships(self, session):
        """Test relationships between benchmark models."""
        # Create a run
        run = BenchmarkRun(
            run_name="Test Run",
            config_hash="test123",
            query_hash_list=["q1"],
            search_config={},
            evaluation_config={},
            datasets_config={},
            status=BenchmarkStatus.COMPLETED,
        )
        session.add(run)
        session.commit()

        # Add multiple results
        for i in range(5):
            result = BenchmarkResult(
                benchmark_run_id=run.id,
                example_id=f"example_{i}",
                query_hash=f"hash_{i}",
                dataset_type=DatasetType.SIMPLEQA,
                question=f"Question {i}",
                correct_answer=f"Answer {i}",
                is_correct=i % 2 == 0,  # Alternate correct/incorrect
            )
            session.add(result)

        session.commit()

        # Test relationship queries
        run_with_results = session.query(BenchmarkRun).first()
        results = run_with_results.results.all()
        assert len(results) == 5

        # Count correct results
        correct_count = run_with_results.results.filter_by(
            is_correct=True
        ).count()
        assert correct_count == 3

    def test_unique_constraints(self, session):
        """Test unique constraints on benchmark models."""
        run = BenchmarkRun(
            config_hash="test123",
            query_hash_list=["q1"],
            search_config={},
            evaluation_config={},
            datasets_config={},
        )
        session.add(run)
        session.commit()

        # Add a result
        result1 = BenchmarkResult(
            benchmark_run_id=run.id,
            example_id="test_001",
            query_hash="unique_hash",
            dataset_type=DatasetType.SIMPLEQA,
            question="Test question",
            correct_answer="Test answer",
        )
        session.add(result1)
        session.commit()

        # Try to add duplicate (same run_id and query_hash)
        result2 = BenchmarkResult(
            benchmark_run_id=run.id,
            example_id="test_002",
            query_hash="unique_hash",  # Same hash
            dataset_type=DatasetType.SIMPLEQA,
            question="Different question",
            correct_answer="Different answer",
        )
        session.add(result2)

        with pytest.raises(IntegrityError):
            session.commit()

    def test_benchmark_error_handling(self, session):
        """Test error tracking in benchmark runs."""
        run = BenchmarkRun(
            config_hash="test123",
            query_hash_list=[],
            search_config={},
            evaluation_config={},
            datasets_config={},
            status=BenchmarkStatus.FAILED,
            error_message="Failed to connect to evaluation model API",
        )
        session.add(run)
        session.commit()

        # Add failed results
        result = BenchmarkResult(
            benchmark_run_id=run.id,
            example_id="failed_001",
            query_hash="hash_failed",
            dataset_type=DatasetType.SIMPLEQA,
            question="What is 2+2?",
            correct_answer="4",
            research_error="Search timeout after 30 seconds",
            evaluation_error="Could not parse model response",
        )
        session.add(result)
        session.commit()

        # Verify error tracking
        failed_run = (
            session.query(BenchmarkRun)
            .filter_by(status=BenchmarkStatus.FAILED)
            .first()
        )
        assert (
            failed_run.error_message
            == "Failed to connect to evaluation model API"
        )

        failed_result = (
            session.query(BenchmarkResult)
            .filter(BenchmarkResult.research_error.isnot(None))
            .first()
        )
        assert "timeout" in failed_result.research_error
        assert failed_result.evaluation_error is not None

    def test_benchmark_statistics(self, session):
        """Test calculating statistics from benchmark results."""
        # Create a completed run
        run = BenchmarkRun(
            config_hash="test123",
            query_hash_list=["q1", "q2", "q3", "q4", "q5"],
            search_config={},
            evaluation_config={},
            datasets_config={},
            status=BenchmarkStatus.COMPLETED,
            total_examples=5,
            completed_examples=5,
        )
        session.add(run)
        session.commit()

        # Add results with varying accuracy
        accuracies = [True, True, False, True, False]  # 3/5 = 60% accuracy

        for i, is_correct in enumerate(accuracies):
            result = BenchmarkResult(
                benchmark_run_id=run.id,
                example_id=f"test_{i}",
                query_hash=f"hash_{i}",
                dataset_type=DatasetType.SIMPLEQA
                if i < 3
                else DatasetType.BROWSECOMP,
                question=f"Question {i}",
                correct_answer=f"Answer {i}",
                is_correct=is_correct,
                processing_time=2.0 + i * 0.5,
            )
            session.add(result)

        session.commit()

        # Calculate statistics
        results = (
            session.query(BenchmarkResult)
            .filter_by(benchmark_run_id=run.id)
            .all()
        )

        correct_count = sum(1 for r in results if r.is_correct)
        accuracy = correct_count / len(results)
        assert accuracy == 0.6

        # Calculate per-dataset accuracy
        simpleqa_results = [
            r for r in results if r.dataset_type == DatasetType.SIMPLEQA
        ]
        simpleqa_accuracy = sum(
            1 for r in simpleqa_results if r.is_correct
        ) / len(simpleqa_results)
        assert simpleqa_accuracy == 2 / 3  # ~0.667

        # Calculate average processing time
        avg_time = sum(r.processing_time for r in results) / len(results)
        assert avg_time == 3.0  # (2.0 + 2.5 + 3.0 + 3.5 + 4.0) / 5